polyglot-ko

Model Description

Polyglot-Ko is a series of large-scale Korean autoregressive language models made by the EleutherAI polyglot team.

The model consists of 24 transformer layers with a model dimension of 2048, and a feedforward dimension of 8192. The model dimension is split into 16 heads, each with a dimension of 128. Rotary Position Embedding (RoPE) is applied to 64 dimensions of each head. The model is trained with a tokenization vocabulary of 30003.

Training data

Polyglot-Ko-1.3B was trained on 863 GB of Korean language data (1.2TB before processing), a large-scale dataset curated by TUNiB. The data collection process has abided by South Korean laws. This dataset was collected for the purpose of training Polyglot-Ko models, so it will not be released for public use.

Furthermore, in order to avoid the model memorizing and generating personally identifiable information (PII) in the training data, we masked out the following sensitive information in the pre-processing stage:

• `` : bank account number
• `` : resident registration number
• `` : phone number

Training procedure

Polyglot-Ko-1.3B was trained on 213 billion tokens over 102,000 steps on 256 A100 GPUs with the GPT-NeoX framework. It was trained as an autoregressive language model, using cross-entropy loss to maximize the likelihood of predicting the next token.

How to use

This model can be easily loaded using the AutoModelForCausalLM class:

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("EleutherAI/polyglot-ko-1.3b")
model = AutoModelForCausalLM.from_pretrained("EleutherAI/polyglot-ko-1.3b")

Evaluation results

We evaluate Polyglot-Ko-1.3B on KOBEST dataset, a benchmark with 5 downstream tasks, against comparable models such as skt/ko-gpt-trinity-1.2B-v0.5, kakaobrain/kogpt and facebook/xglm-7.5B, using the prompts provided in the paper.

The following tables show the results when the number of few-shot examples differ. You can reproduce these results using the polyglot branch of lm-evaluation-harness and the following scripts. For a fair comparison, all models were run under the same conditions and using the same prompts. In the tables, n refers to the number of few-shot examples.

In case of WiC dataset, all models show random performance.

python main.py \
   --model gpt2 \
   --model